# Web search may use one or two GPUs per ten inference GPUs

Markdown copy of https://realalpha.app/s/web-search-may-use-one-or-two-gpus-per-ten-inference-gpus/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Parag Agrawal](https://realalpha.app/builders/parag-agrawal/) · Parag Agrawal is founder and CEO of Parallel Web Systems. · Founded Parallel Web Systems, which builds agentic web search for AI agents; former Twitter CEO.
- Host: [Molly O'Shea](https://realalpha.app/people/molly-oshea/) · Founder, Sourcery · Ex-Upfront Ventures investor; has interviewed Alex Karp and Palmer Luckey.
- Show: Sourcery
- Episode: [AI Agents Could Drive 10,000x More Web Activity | Former Twitter CEO](https://realalpha.app/v/iRI6gTe1wGM/), Oct 24, 2026
- Moment: 11:42 to 14:46 of the episode (3 min), playing on https://realalpha.app/s/web-search-may-use-one-or-two-gpus-per-ten-inference-gpus/; the episode from there: https://realalpha.app/v/iRI6gTe1wGM/?t=702

## Quote

> “for every 10 gpus the world spends on model inference will spend **one to two on web search**”
> — [Parag Agrawal](https://realalpha.app/builders/parag-agrawal/) at [14:16](https://realalpha.app/v/iRI6gTe1wGM/?t=856) · Sourcery

## Takeaways

1. He says the broad direction of model scaling was predictable.
2. At fixed capability, models should keep getting smaller and cheaper.
3. He expects one or two web-search GPUs per ten inference GPUs.

## About this segment

Parag Agrawal says his agent thesis has followed a simple scaling trend: better, longer-running models get cheaper at fixed capability. He expects web search to take 10–20% as much GPU capacity as inference.

From “AI Agents Could Drive 10,000x More Web Activity | Former Twitter CEO”, Oct 24, 2026. [Watch the full episode](https://realalpha.app/v/iRI6gTe1wGM/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[11:55](https://realalpha.app/v/iRI6gTe1wGM/?t=715) you've been obsessed with this for a while. So what does that mean and how? So I think I got obsessed with it now almost three and a half years ago. I was the, I don't know if people remember, there was like this moment of this baby AGI and auto GPT speaking of oftentimes developers in the open source world being ahead of everything else. People were building with like, I don't know what is the model of the day, GP35 or whatever, like old models that weren't, like relative to today, they were like really bad models, relative to today.

[12:28](https://realalpha.app/v/iRI6gTe1wGM/?t=748) And people were trying all these things and I was trying all of these things to try and make things work. And they didn't work. But if you believed you spent any time with the labs back then, you internalized how models will scale and how models will improve like in some sense today it feels like everything changes every month in some sense i don't know if it's every month it might be every week but if you look at it a different way almost nothing that is happening is diverging from what you could have predicted in a straight line three years ago yeah models will keep getting bigger and better. They will run for longer. Holding capability constant, models will get way way way smaller and cheaper. And people will be able to productize around them to disrupt every category of work that can be done on a computer. These were very

[13:30](https://realalpha.app/v/iRI6gTe1wGM/?t=810) simple things that we believed and said three years ago. And we have only seen evidence week on week and month on month supporting these very few simple things. And almost everything follows. So now you can say, yeah, I didn't predict that the big category would be coding agents first. I didn't predict the CLI form factor. Those were all bits of information in news to me along the way. But the macro thesis of agents are going to get more capable.

[14:07](https://realalpha.app/v/iRI6gTe1wGM/?t=847) You're going to do a bunch of inference. And for all the inference you do, you're going to have to do web search along it. I've now been on the record everywhere saying like the one invariant I believe is we will use a lot of GPUs for model inference for agents, whether they be at work or consumer agents and for every 10 gpus the world spends on model inference will spend one to two on web search